Writing
Latest from the journal
- apple-silicondtracemacos
Reading Apple Silicon's performance counters without root
There is no eBPF for macOS and DTrace requires root. This paper documents what remains: sample, xctrace, the undocumented kpep event catalogue, and a method for identifying the unnamed counter slots that xctrace exports.
- analyticsdata-engineeringdata-quality
Data Quality Is Not About Pristine Data. It Is About Being Fit for Use.
A practical framework for measuring data quality, understanding its costs, and focusing on what actually matters for the people who use the data.
- bundrizzlenodejs
Postgres connection options you might not know
A grab bag of Postgres connection/session defaults (and one Linux I/O gotcha) that make your app safer, more predictable, and easier to debug.
Work
Selected projects
01
turbofile
Real async file I/O for Python. Existing async-file libraries hop every call through a thread pool. This one submits the read or write to the kernel, io_uring on Linux and POSIX AIO on macOS, and completes your await when the data has moved. Zero-copy on the hot paths, one event-loop wakeup per burst of completions, aiofiles-compatible API. Repo: https://github.com/4thel00z/turbofile
RustPythonPyO3io_uringPOSIX AIO02
pdfboss
From-scratch PDF toolkit in Rust with Python bindings. The fastest PDF library in the world for page rendering and text extraction, and it creates PDFs too.
RustPythonPyO3PDF03
kaktus-bbox-lora-v3
LoRA adapter for Qwen3-VL-4B that reads a rendered document page and returns its layout as JSON: element types, bounding boxes, text. Trained on FinePDFs pages rendered with pdfboss.
LoRAQwen3-VLDocument AIms-swift04
aria-embedding-311m-en-de
English and German retrieval embedding model finetuned from IBM Granite. 32k context with Matryoshka embeddings, strongest on long documents and German legal retrieval.
sentence-transformersModernBERTMatryoshkaRetrieval05
llmns
A URI scheme for language models. One llm:// reference names the host, protocol, model version and credential. Rust core with Python bindings.
RustPythonPyO3URI schemeRFC06
pdfarena
In-browser benchmark that races four PDF engines on a document you supply: pdfboss, hayro, pdf.js and PDFium, each in its own web worker. The stopwatch wraps only the render call, and the full sweep runs one engine at a time so they cannot fight over cores. Nothing is uploaded.
RustWebAssemblyWeb WorkersNuxtVue
Newsletter
Get new posts by email
One email when something new is published. Unsubscribe anytime.